Pareto optimal, strategy proof, and non-bossy
matching mechanisms
Sophie Bade∗
November 14, 2014
Abstract
The set of all Pareto optimal, strategy proof and non-bossy mecha-
nisms is characterized as the set of trading and braiding mechanisms.
Fix a matching problem with more than three houses and a profile
of preferences. At the start of a trading and braiding mechanism at
most one house is brokered; all other houses are owned. In the first
trading round, owners point to their most preferred houses, the bro-
ker - if there is one - points to his most preferred owned house, and
houses point to the agents who control them. At least one cycle forms.
The agents in such a cycle are matched to the houses they point to.
The process is repeated with the remainder. Once there are only three
houses left the mechanism might turn into a braid, a device that avoids
a particular matching.
Keywords: . JEL Classification Numbers: C78.
∗Royal Holloway College, University of London and Max Planck Institute for Research
on Collective Goods, Bonn [email protected]. Funding from the ARCHES prize is
gratefully acknowledged.
1
1 Introduction
I characterize the set of all Pareto optimal, strategy proof, and non-bossy
matching mechanisms. There are finitely many agents and objects, called
houses. Agents strictly rank any two different houses and prefer any house
to homelessness. Mechanisms map profiles of preferences to matchings. A
matching maps agents to houses in such a way that each agent obtains at
most one house and no house is matched to two different agents. A mech-
anism is Pareto optimal if it maps any profile of preferences to a Pareto
optimal matching. It is strategy proof if no agent can improve his match by
submitting a false preference. In a non-bossy mechanism, an agent’s change
of preference can change some other agent’s match only if it also changes
the agent’s own match. I call a mechanism that satisfies these three criteria
good.
The main result of the paper (Theorem 1) is that a mechanism is good if
and only if it is a trading and braiding mechanism.
At the start of a trading and braiding mechanism all houses are either
owned or brokered. An owner might own multiple houses. There is at most
one broker and he brokers at most one house. Matchings are determined
through a sequence of trading cycles. To construct cycles, any house points
to the agent who owns or brokers it. Any owner points to his most preferred
house. The broker, if there is one, points to his most preferred house among
the owned ones. (A broker may not point to the house he brokers.) At least
one cycle forms. Each agent in such a cycle is matched with the house he is
pointing to. All matched agents and houses exit the mechanism.
A new round of pointing ensues. A fixed rule determines the ownership
and brokerage rights over all unmatched houses. This rule is such that any
owner who remains unmatched continues to own (at least) the houses he
owned at the start of the preceding round. Moreover any unmatched broker
continues to broker the same house if at least two owners from the preceding
round remain unmatched. As long as at least four houses remain unmatched,
the mechanism continues as described so far.
If following some round there are exactly three houses left, the mechanism
may turn into a braid. A braid is defined via an arbitrary matching that I
2
call an avoidance-matching. For any profile of preferences the braid selects a
Pareto optimum that minimally coincides with the avoidance matching. The
name braid derives from the intricate and symmetric relationship between
all three agent-house pairs (strands) in the avoidance matching. The trading
and braiding process terminates once a matching is reached.
A mechanism is group-strategyproof if no group of agents can misstate
their preferences to make at least one of its members better off while making
no member worse off. Combining Papai’s [14] result that a matching mech-
anism is group-strategyproof if and only if it is non-bossy and strategyproof
with Theorem 1 it follows that the set of group-strategyproof and Pareto op-
timal mechanisms coincides with the set trading and braiding mechanisms.
Although the current paper is most closely related to Pycia and Unver [16]
let me first review the long tradition that has studied good mechanisms. This
literature starts with the definition of two canonical mechanisms. According
to serial dictatorship one agent, the first dictator, gets to choose a house out
of the grand set. Then another agent, the second dictator, gets to choose
out of the remainder and so forth. A second mechanism, Gale’s top trading
cycles (Shapley and Scarf [18]), applies to the case of equally many agents
and houses: each agent starts out owning exactly one house and points to
his most preferred house. At least one cycle forms. All agents in all such
cycles are matched with the houses they point to. The procedure is repeated
until a matching is obtained. Svensson [19] characterized serial dictatorship
as the unique mechanism that satisfies strategy proofness, non-bossiness and
neutrality in the sense that the outcome of the mechanism does not depend
on the names of the houses. Ma [13] characterized Gale’s top trading cycles
as the unique Pareto optimal and strategy proof mechanism that satisfies
individual rationality in the sense that every agent is matched to a house
that he likes at least as much as the house he is endowed with. Svensson
[19] does not not impose Pareto optimality, whereas Ma [13] does not impose
non-bossiness. In the two characterizations, these two properties arise as a
consequence of the respective other three criteria imposed.
Papai [14] showed a mechanism is good and reallocation proof if and only
if it can be represented as a hierarchical exchange mechanism. A mecha-
nism is reallocation proof if no two individuals can gain by misrepresenting
3
preferences and then swapping the houses they are matched to under the
mechanism. The difference between Gale’s top trading cycles mechanisms
and hierarchical exchange mechanisms lies in the fact that agents may own
multiple houses in hierarchical exchange mechanisms. The set of hierarchical
exchange mechanisms coincides with the set of trading and braiding mech-
anisms without brokers or braids.1 Brokers entered the matching literature
with Pycia and Unver [16]’s trading cycles mechanisms.
The present paper would not have been possible without the work by
Papai [14] and by Pycia and Unver [16]. My proof directly employs a Lemma
from Papai [14] and various steps of my proof mirror similar chapters in the
proofs of Papai [14] and of Pycia and Unver [16]. More importantly my
characterization relies on the basic and ingenious ideas of constructing new
good mechanisms by extending Gale’s top trading cycles mechanisms to allow
for the ownership of multiple houses (Papai [14]) and brokerage (Pycia and
Unver [16]).
Pycia and Unver [16] set out with the same goal as the current paper.
They defined trading cycles mechanisms with the aim to to characterize the
set of all good mechanisms. However, the characterization by Pycia and
Unver [16] is not correct as they overlooked braids. Braids are good - but
they are not representable in the framework of Pycia and Unver [16]. So the
class of Pycia and Unver’s [16] trading cycles mechanisms is strictly nested
between Papai’s [14] hierarchical exchange mechanisms and the set of all
good mechanisms.
A variety of papers investigates how the grand set of all good mecha-
nisms is restricted by some additional assumptions. Che, Kim and Kojima
[8] for example contrast the non-existence of an ex post incentive compati-
1Abdulkadiroglu and Sonmez [3] defined the set of you-request-my-house-I-get-your-
turn mechanisms. There mechanisms have been characterized by Sonmez and Unver [17]
as the set of all Pareto optimal, strategy proof, individually rational, weakly neutral and
consistent mechanisms. Abdulkadiroglu and Sonmez [4] defined the set of all top trading
cycles mechanisms, which were then characterized by Abdulkadiroglu and Che [1]. Serial
dictatorships and Gale’s top trading cycles mechanism are both you-request-my-house-I-
get-your-turn mechanisms. The set of top trading cycles mechanisms is strictly nested
between the sets of you-request-my-house-I-get-your-turn mechanisms and hierarchical
exchange mechanisms.
4
ble and Pareto optimal mechanism when agents have interdependent values
with the large set of Pycia and Unver’s trading cycles mechanisms. Velez [20]
uses trading cycles mechanisms to put into relief the restrictions that arise
when one imposes consistency upon good mechanisms. Ehlers and West-
kamp [9] highlight the limited scope for strategyproofness when allowing for
indifferences by a comparison with the large set trading cycles mechanism
- all of which are strategyproof under the assumption that preferences are
linear orders. In some cases, knowing the grand set, convoluted as it may
be, allows us to draw stronger conclusions about simple mechanisms. Bade
[6], for example, argues that the uniform randomization of any good mecha-
nism results in a random serial dictatorship. The proof of this result heavily
relies on the characterization of the set of all good mechanisms. In this
case, the characterization of all good mechanisms strengthens the appeal of
one - very simple - mechanism: random serial dictatorship. Similarly Bade
[5] claims that with endogenous information acquisition there is exactly one
ex ante Pareto optimal, strategy proof, and non-bossy mechanism: serial
dictatorship. Once again the proof starts by considering the grand set of
all mechanisms that satisfy these three criteria under the assumption that
information is exogenously given.
The characterization of the grand set of all good mechanisms allow us to
precisely identify goals that are achievable in addition to Pareto optimality,
strategyproofness and non-bossiness. Braids and brokerage are design tools
to avoid certain matchings. Suppose that a designer wishes to use a good
mechanism to match a set of administrators to a set of tasks via good mech-
anism with the additional caveat that some secretary, say Amy, should not
proctor exams. By letting Amy broker the task of proctoring, the designer
ensures that she will only be matched to this task if all other secretaries pre-
fer their respective matches to proctoring. Similarly if there are exactly three
tasks and three secretaries there exists a good mechanism that allows the de-
signer to specify some matching which is to be avoided: the designer can use
a braid. While trading and braiding mechanisms map out a clear path as to
how to incorporate avoidance goals into good mechanism, they at the same
time show that not many of such goals can be accommodated simultaneously,
given that braids are only defined when there are exactly three houses and
5
given that there is at most one brokered house in any trading round that is
not a braid.
2 Concepts
A housing problem consists of a set of agents N : = {1, · · · , n}, a finite
set of houses H and a profile of preferences R = (Ri)ni=1. The option to
stay homeless ∅ is always available: ∅ ∈ H. An agent’s preference Ri is
a linear order2 on H and each agent prefers any house to homelessness, so
hRi∅ holds for all i, h ∈ H. The set of all profiles of preferences is R.
The notation hRiH′ means that agent i prefers h to each house in H ′. If
an agent i holds a preference eRigRiH \ {e, g}, I write Ri : e g. For any
fixed profile R, the preferences of a group of agents G ⊂ N is denoted RG,
similarly R−G denotes the preferences of all agents not in this group, so we
have R = (RG, R−G) = (Ri, R−i). The profile R is the restriction of R
to some N ′ ⊂ N and H ′ ⊂ H if eRig ⇔ eRig for g, h ∈ H ′ and i ∈ N ′.
Two preferences Ri, R′i coincide on H ′ ⊂ H if eRig ⇔ eR′ig holds for all
e, g ∈ H ′. So Ri, R′i coincide on H ′ if and only if their restrictions to H ′ are
identical.
Submatchings match subsets of agents to at most one house each. A
submatching is a function ν : N → H such that ν(i) = ν(j) and i 6= j
imply ν(i) = ∅. The sets of agents and houses matched under ν are Nν : =
N \ ν−1(∅) and Hν : = ν(Nν). When ν(i) 6= ∅ then ν(i) is the house that
agent i is matched to under ν; N ν : = N \ Nν and Hν : = H \ Hν are the
sets of agents and houses not matched by ν. When convenient I interpret a
submatching ν as the set of agent-house pairs {(i, h) : ν(i) = h 6= ∅}. For
two submatchings ν and ν ′ with Nν ∩Nν′ = ∅ = Hν ∩Hν′ the submatching
ν ∪ ν ′ : Nν ∪ Nν′ → Hν ∪ Hν′ is defined by (ν ∪ ν ′)(i) = ν(i) if i ∈ Nν and
(ν ∪ ν ′)(i) = ν ′(i) otherwise. A submatching µ is a matching if Hµ = H or
Nµ = N (or both) hold, the set of all matchings is M.
A mechanism is a function M : R → M. The outcome of M at R,
M(R), matches agent i to house M(R)(i). A mechanism M is Pareto opti-
2So hRih′ and h′Rih together imply h = h′.
6
mal if for no R there exists a matching µ′ 6= M(R) such that µ′(i)RiM(R)(i)
for all i.3 A mechanism M is strategy proof if M(R)(i)RiM(R′i, R−i)(i)
holds for all R ∈ R, all R′i and all agents i: declaring one’s true preferences is
a weakly dominant strategy in a strategy proof mechanism. A mechanism M
is non-bossy if M(R)(i) = M(R′i, R−i)(i) implies M(R)(j) = M(R′i, R−i)(j)
for all R,R′i and all i, j ∈ N , so an agent can only change someone else’s
match if he also changes his own match. A Pareto optimal, strategy proof
and non-bossy mechanism is a good.
3 Braids
A braid B : R → M is a mechanism for a problem with exactly three
houses and at least as many agents. It is fully defined through an avoidance
matching ω. Matchings B(R) are chosen to avoid matching i to ω(i) while
keeping the set of matched agents equal to the set of agents matched under
ω. For any R let PO(R) be the set of Pareto optima µ with Nω = Nµ. If
minµ∈PO(R) | {i : µ(i) = ω(i)} | is attained at a unique µ∗ then let B(R) = µ∗.
If not, at least two agents in Nω must rank some house h∗ = ω(i∗) at the
top and the pair (i∗, h∗) is decisive in the following sense. If only one agent
j 6= i∗ ranks h∗ at the top then B(R) is the unique minimizer that matches
j to h∗. If both agents i 6= i∗ rank h∗ at the top, then B(R) is the unique
minimizer preferred by i∗.
To concretely illustrate braids let H = {e, g, g′} and | N |≥ 3. Since any
agent i stays unmatched under the avoidance matching (ω(i) = ∅) if and
only if he stays unmatched at any outcome of the braid (B(R)(i) = ∅ for
all R) and since B is non-bossy it is w.l.o.g to assume that N = {1, 2, 3}.Given | H |=| N |= 3 it is convenient to denote matchings as vectors with
the understanding that the i-th component of such a vector represents agent
i’s match. Moreover given that there are only three agents, the requirement
Nω = Nµ for any µ = B(R) is automatically satisfied and can therefore be
ignored. Let ω : = (e, g′, g). There are exactly two matchings ω′ 6= ω′′ with
ω′(i) 6= ω(i) 6= ω′′(i) for all i. Let ω′ : = (g, e, g′), and ω′′ : = (g′, g, e). If
3Since all Ri are linear at least one agent must strictly prefer µ′(i) to M(R)(i) if
µ′ 6= M(R).
7
Ri : e g for i = 1, 2, 3 then minµ∈PO(R) | {i : ω(i) = µ(i)} | is attained both
at ω′ and ω′′. Since at least two agents rank house e = ω(1) at the top,
agent 1 is decisive. Agent 1’s preference of g over g′ implies B(R)(1) = g and
B(R) = ω′. Under (R′2, R−2) with R′2 : g g′ (and Ri as above) there are four
Pareto optima: ω, (e, g, g′), (g, g′, e) and ω′′. Since ω′′ is the unique Pareto
optimum with | {i : ω(i) = µ(i)} |= 0, B(R) equals ω′′.
Lemma 1 The braid B is good.
Proof W.l.o.g fix ω, ω′ and ω′′ as in the preceding paragraph and assume
that N = {1, 2, 3}.
(*) If R is such that for some j 6= j′, Rj : ω′(j), ω′(j′)Rj′ω′′(j′) and ω(j′) =
ω′(j) hold, then B(R) = ω′. To prove (*) assume w.l.o.g that j = 2, so
ω′j(2) = e, ω(1) = e, j′ = 1 and ω′(1) = g. Since ω′ is Pareto optimal at R
B(R) = ω′ holds if ω′′ /∈ PO(R). If ω′′ ∈ PO(R) then R3 : e has to hold. But
then agent 1 is decisive and B(R) = ω′ holds since g = ω′(1)R1ω′′(1) = g′
(**) If B(R)(i) = ω(i), then i is the only agent who under R ranks ω(i) at
the top.
Any profile R is mapped to a Pareto optimum according to B(R), so B
is Pareto optimal.
To see that B is strategy proof fix a profile R an agent i and a deviation
R′i. If there is a unique Pareto optimum at R each agent obtains his most
preferred house and we have B(R)(i)RiB(R′i, R−i). Now consider the case
that at least two agents rank the same house at the top. W.l.o.g let this
house be e = ω(1).
Case 1: gR1g′ and R2 : e. Then (*) implies B(R) = ω′ and (**) im-
plies that g is the best attainable house for agent 1 given R−1. Agent 2
is matched to his most preferred house under B(R). Finally (*) implies
B(R) = B(R′3, R−3) for all R′3. So B(R)(i)RiB(R′i, R−i) holds. Mutatis mu-
tandis, the same reasoning implies that B(R)(i)RiB(R′i, R−i) also holds in
Case 2: g′R1g and R3 : e.
Case 3: gR1g′ but not R2 : e. Since at least two agents rank e at
the top R1 : e g and R3 : e must hold. If R2 : g′, then (g, g′, e) ∈PO(R) ⊂ {ω, (e, g, g′), (g, g′, e)}, where the latter two matchings minimize
8
| {i : ω(i) = µ(i)} |. Observation (**) implies B(R) = (g, g′, e). Since
agents 2 and 3 obtain their most preferred house, they have no incentive to
deviate. Moreover Observation (**) implies that B(R′1, R−1)(1) 6= e, and
consequently agent 1 has no incentive to deviate. If R2 : g, ω′′ ∈ PO(R) but
ω′ /∈ PO(R), so B(R) = ω′′. Once again agents 2 and 3 obtain their most
preferred house and have no incentive to deviate. Observation (**) implies
that B(R′1, R−1)(1) 6= e. On the one hand we have ω′′ ∈ PO(R′1, R−1) for all
R′1. On the other hand ω′ is only an element of PO(R′1, R−1) if R1 : g. But
then agent 3 is decisive as ω(g) = 3. Since g′R3e B(R′1, R−1)(3) = g′ and
therefore B(R′1, R−1) = ω′. So B(R)(i)RiB(R′i, R−i) holds when gR1g′ but
not R2 : e. The same arguments apply mutatis mutandis to Case 4: g′R1g
but not R3 : e. In sum B is strategy proof.
To see that B is non-bossy, let B(R)(1) = B(R′1, R−1)(1) : = h, so
B(R) = {(1, h)} ∪ ν∗ and B(R′1, R−1) = {(1, h)} ∪ ν◦ for some submatchings
ν∗, ν◦. Since B is Pareto optimal ν∗, ν◦ ∈ PO(R) where R is the restriction
of agent 2 and 3’s preference to H \{h}. If PO(R) is a singleton then ν∗ = ν◦
and we are done. So suppose PO(R) is not a singleton. If h = ω(1), then
PO(R) = {ν, ν ′} where ν(2) = ω(2) and ν ′(2) = ω(3). So {(1, h)} ∪ ν = ω
and therefore minµ∈PO(R) | {i : µ(i) 6= ω(i)} | can only be attained at
{(1, h)} ∪ ν ′, implying ν ′ = ν∗ = ν◦. If h 6= ω(1), then PO(R) = {ν, ν ′}is such that {(1, h)} ∪ ν ′ ∈ {ω′, ω′′} and {(1, h)} ∪ ν /∈ {ω′, ω′′}. Once
again minµ∈PO(R) | {i : µ(i) 6= ω(i)} | can only be attained at {(1, h)} ∪ ν ′,implying ν ′ = ν∗ = ν◦. Having covered all cases we can conclude that
B(R)(1) = B(R′1, R−1)(1) implies B(R) = B(R′1, R−1) and B is non-bossy.
�
4 Trading and braiding mechanisms
A control rights function at some submatching ν cν : Hν → Nν × {o, b}assigns control rights over any unmatched house to some unmatched agent
and specifies a type of control. If cν(h) = (i, x), then agent i controls house
h at ν. If x = o, then i owns h; if x = b he brokers h. Control rights
functions satisfy the following three criteria:
9
(C1) If more than one house is brokered, then there are exactly three houses
and they are brokered by three different agents.
(C2) If exactly one house is brokered then there are at least two owners.
(C3) No broker owns a house.
A general control rights structure c maps a set of submatchings ν
to control rights functions cν . For now just assume that c is defined for
sufficiently many submatchings to ensure that the following algorithm is well
defined for any fixed R.
Initialize with r = 1, ν1 = ∅
Round r: only consider the remaining houses and agents Hνr and Nνr .
Braiding: If more than one house is brokered under cνr let B be the braid
defined by the avoidance matching ω with cνr(ω(i)) = (i, b). Terminate the
process with M(R) = νr ∪ B(R) where R is the restriction of R to Hνr and
N νr . If not, go on to the next step.
Pointing: Each house points to the agent who controls it, so h ∈ Hνr points
to i ∈ N νr with cνr(h) = (i, ·). Each owner points to his most preferred house,
so owner i ∈ Nνr points to house h ∈ Hνr if hRiHνr . Each broker points
to his most preferred owned house, so broker j ∈ N νr with cνr(hb) = (j, b)
points to house h ∈ Hνr \ {hb} if hRjHνr \ {hb}.
Cycles: Select at least one cycle. Define ν◦ such that ν◦(i) : = h if i points
to h in one of the selected cycles.
Continuation: Define νr+1 : = νr ∪ ν◦. If νr+1 is a matching terminate the
process with M(R) = νr+1. If not, continue with round r + 1.
The trading and braiding process starts with the initial submatching ν1 =
∅. If c does not call for a braid at ν1, each house points to the agent who
controls it according to cν1 . Each owner points to his most preferred house,
and the only broker (if there is one) points to his most preferred house among
the owned houses. At least one cycle forms. All agents and houses in some
chosen cycles are matched, yielding the submatching ν2. A new round of
10
either braiding or pointing cycles ensues. Once a matching is reached the
process terminates.
A submatching ν is reachable under c at R if some round of a trading
and braiding process can start with ν. A submatching ν is c-relevant if
it is reachable under under c at some R.4 A submatching ν ′ is a direct
c-successor of of some c-relevant ν if there exists a profile of preferences R
such that ν is reachable under c(R) and ν ′ arises out of matching a single
cycle at ν. A control rights structure c maps any c-relevant submatching
ν to a control rights function cν and satisfies requirements (C4), (C5), and
(C6).
Fix a c-relevant submatching ν◦ together with a direct c-successor ν.
(C4) If i /∈ Nν owns h at ν◦ then i owns h at ν.
(C5) If at least two owners at ν◦ remain unmatched at ν and if i /∈ Nν brokers
h at ν◦ then i brokers h at ν.
(C6) If i owns h at ν◦ and ν and if i′ /∈ Nν brokers h′ at ν◦ but not at ν,
then i owns h′ at ν and i′ owns h at ν ∪ {(i, h′)}.
In a braid no agent owns any house. Consequently, c can only map a
submatching ν to a braid (a control right function cν with three brokered
houses), if (C4), (C5) and (C6) do not require any ownership at ν. The
trading and braiding mechanism defined by the control rights structure c is
also denoted c and c(R) is the outcome of the trading and braiding mechanism
c at the profile of preferences R.
Any c-relevant submatching ν defines a submechanism c[ν] that maps
restrictions R (of some R ∈ R to Nν and Hν to submatchings) µ′ with the
feature that ν ∪ µ′ is a matching in the original problem. The control rights
structure defining c[ν] is such that ν◦ = ν ∪ ν ′ is c-relevant if and only if ν ′
4Consider a control rights structure c with three agents {1, 2, 3} and 4 houses
{h, g, k, h′}, where agent 1 starts out owning house h and g and agent 2 starts out owning
the remaining houses. Suppose the profile of preferences R is such that 1 most prefers
h, and 2 most prefers h′. Then the following submatchings involving agents 1 and 2 are
reachable under c(R): {(1, h)}, {(2, h′)} and {(1, h), (2, h′)}. The submatching {(1, g)} is
c-relevant since agent 1 could appropriate house g, but not it is not reachable under c(R)
given that 1 prefers h to g. The submatching {(3, h)} is not c-relevant since 3 does not
own any house at the start of the mechanism.
11
is c[ν]-relevant. For any such pair ν◦, ν ′ we have c[ν]ν′ = cν◦ . It is easy to
check that c[ν] also defines a trading and braiding mechanism. Fixing any
c, R, and ν that is reachable under c(R), the definition of the trading-cycles
process implies c(R) = ν ∪ c[ν](R), where R is the restriction of the profile
of preferences R to the set of agents Nν and the houses Hν .
5 Theorem
Theorem 1 Any good mechanism has a unique representation as a trading
and braiding mechanism. Any trading and braiding mechanism is good.
A trading and braiding mechanism is a hierarchical exchange mechanism
following Papai [14] if there are no brokers or braids. Just as in hierarchical
exchange mechanisms, owners in trading and braiding mechanisms are free
to either appropriate or trade any house he owns. A trading a braiding cycles
mechanism is a trading cycles mechanism following Pycia and Unver [16] if
it does not have any braids. So Pycia and Unver’s [16] notion of brokerage
is identical to the present one. A broker may not appropriate the house he
brokers. He is however free to exchange this house as he pleases. Moreover
when ν is not mapped to a braid there is at most one broker at any given
round and he brokers at most one house.
There is one important formal difference between the definition of trading
and braiding mechanisms and its predecessors: Trading and braiding mech-
anisms only require that at least one cycle is matched in any one round. All
other definitions based on sequences of trading cycles, that I am aware of,
require the matching of all cycles at any given round. The disadvantage of
the current definition is that it is a priori not clear that all different orders
of elimination of trading cycles lead to the same outcome. I therefore need
to show that trading and braiding mechanisms are well-defined (see Section
6.1). But once this proof is out of the way the proofs that trading and
braiding mechanisms are strategy proof and non-bossy turn out to be much
12
easier.5
Pycia and Unver [16] introduced the notion of control rights functions
into the matching literature. The approach rendered the definition of any
mechanism that use sequences of trading rounds significantly simpler. With
my definition of trading and braiding mechanisms I follow their lead. How-
ever, I only require that control rights functions cν be defined on c-relevant
submatchings while Pycia and Unver’s [16] control rights structures are de-
fined on all submatchings that are not themselves matchings. Moreover the
present requirement (C2) that there are at least two owners when there is a
broker, strengthens Pycia and Unver’s [16] requirement (R2) that an agent
must own all unmatched houses if he is the only unmatched agent. The latter
two differences allow for the uniqueness statement in Theorem 1. If one was
to replace (C2) as stated here with Pycia and Unver’s [16] (R2) the class of
mechanisms described would not change. However the uniqueness result in
Theorem 1 would no longer hold.
6 Proof
I prove Theorem 1 by induction over the number of agents n.
Start: If n = 2 then any good mechanism is either a serial dictatorship or
Gale’s top trading cycles mechanism, both of which are trading and braiding
mechanisms.
Suppose that trading and braiding mechanisms with n agents are well-
defined. Suppose that any good mechanism with n agents can be represented
as a unique trading and braiding mechanism and that any trading and braid-
ing mechanism with n agents is good. Fix a control rights structure c with
5Bade’s [6] proof that the symmetrization of any good mechanism is equivalent to
random serial dictatorship relies on the flexibility to eliminate trading cycles in any which
order. Carroll [7] also uses this flexibility in his proof that the symmetrization of any top
trading cycles mechanism is equivalent to random serial dictatorship. Since the definition
of top trading cycles mechanism requires the immediate elimination of all trading cycles
at any stage, Carroll [7] provides a proof of the equivalence of all orders of elimination of
trading cycles in top trading cycles mechanisms.
13
n+1 agents that is not itself a braid and a profile of preferences R. In Section
6.1 I show that the order of the elimination of trading cycles does not matter
for the outcome c(R), so c indeed defines a mechanism. Section 6.2 shows
that c is good. Section 6.3 shows that the representation is unique.
Section 6.4 lists a set of arguments that are repeatedly used in the up-
coming proof that any good mechanism M can be represented as a trading
and braiding mechanism. For M to be representable as a trading and braid-
ing mechanism c one needs to pinpoint for any house an agent who controls
this house at ∅. Here I follow the proof of Pycia and Unver [16] and define
a function c∅ : H → N × {o, b} in Section 6.5. Section 6.6 shows that this
function c∅ shares a set of useful properties with control rights functions as
defined in Section 4. Section 6.7 shows that braids are the only alternative
to trading rounds, so c∅ satisfies (C1). In the same section I show that c∅also satisfies (C2) and (C3), implying that c∅ is indeed a control rights func-
tion. Section 6.8 shows that the outcome of M(R) when M is not a braid
is consistent with the submatching achieved in a first trading round under
c∅ at that R. The inductive hypothesis is used to show that any submatch-
ing that can be reached via the formation of a single cycle at c∅ under any
R is followed by a well-defined trading and braiding (sub-)mechanism c[ν].
These submechanisms c[ν] together with c∅ define a control rights structure
c. Section 6.9 finishes the proof by showing that c satisfies the remaining
requirements (C4), (C5) and (C6) that link control rights between rounds.
6.1 Trading and braiding mechanisms are well-defined
Fix a control rights structure c for n + 1 agents and a profile of preferences
R. To see that the order of the elimination of trading cycles does not matter
let νs and νa arise out of matching respectively an arbitrary (non-empty)
subset or all of the cycles at ∅. It suffices to show that νa is reachable after
the agents and houses in some chosen cycles are matched to obtain νs. The
hypothesis of the induction implies that the order of elimination does not
matter in c[νa]. If there is exactly one cycle at ∅, νs = νa must hold and we
are done. So suppose there are at least two cycles at ∅ and let νs ( νa. Since
no broker may point to the house he brokers at least one agent in Nνa \Nνs is
14
an owner at ∅. Given (C4) this agent must also be an owner at νs. Therefore
c cannot map νs to a braid.
Case I: c∅(h) = cνs(h) holds for all h ∈ Hνa \Hνs . So any h ∈ Hνa \Hνs
points to the same agent at ∅ and at νs. Moreover any agent i ∈ Nνa \ Nνs
points to the same house at ∅ and at νs (since h∗RiH implies h∗RiHνs and
since h∗RiH \ {hb} implies h∗RiHν \ {hb}, which is relevant if i brokers hb).
So any cycle at ∅ that is not immediately eliminated remains a cycle at νs.
By the inductive hypothesis the order of the elimination of cycles does not
matter in the submechanism c[νs]. We can start by matching all cycles that
already existed at ∅ and νa is reachable.
Case II: c∅(h) 6= cνs(h) holds for some h ∈ Hνa \Hνs . By (C4) h cannot
be owned at ∅ so c∅(h) = (ib, b) holds for some ib. As a broker ib must point
to some owned house h∗ at ∅. Let c∅(h∗) = (i∗, o). By (C4) agent i∗ also
own h∗ at νs. By (C5) i∗ is the only agent who owns houses at ∅ and νs. In
sum, the cycle ib → h∗ → i∗ → hb forms at ∅ and νa = νs ∪ {(i∗, h), (ib, h∗)}.
By (C6) i∗ owns h at νs. Since i∗ points to h at ∅ we have hRi∗H and
consequently hRi∗Hνs . So at νs, h points to i∗ and i∗ to h. By the inductive
hypothesis the order of the elimination of cycles does not matter in c[νs], so
νs ∪ {(i∗, h)} is reachable. (C6) implies that at νs ∪ {(i∗, h)} agent ib owns
all houses owned by i∗ at ∅, in particular house h∗. Given that h∗RibH \ {h}implies h∗RibHνs\{h} a cycle just involving ib and h∗ forms. By the inductive
hypothesis we may eliminate this one cycle and νa = νs ∪ {(i∗, h), (ib, h∗)} is
reachable.
6.2 Trading and braiding mechanisms are good
Let N (c∅) be the the set of direct c-successors to ∅ and let N (c∅)(R) be
the subset of direct c-successors to ∅ that are reachable under c(R). Since
these sets coincide for any two different trading and braiding mechanisms
c, c′ with c∅ = c′∅ only the control rights function c∅ is used to define N (c∅)
and N (c∅)(R). Fix an agent i and a preference R′i. For any ν ∈ N (c∅)
let R and R′i be the restrictions of R and R′i to N ν , Hν . If ν ∈ N (c∅)(R)
the definition of the trading process implies that c(R) = ν ∪ c[ν](R) and
c(R′i, R−i) = ν ∪ c[ν](R′i, R−i) if i /∈ Nν . The proof that c is strategyproof
15
and non-bossy of c is split into two cases.
Case I: There exists some ν ∈ N (c∅)(R) with i /∈ Nν . Fix such a ν.
Since c[ν] involves at most n agents, it is strategy proof by the inductive
hypothesis and c[ν](R)(i)Ric[ν](R′i, R−i)(i) holds. Given that Ri and Ri co-
incide on Hν , c[ν](R)(i) = c(R)(i), and c[ν](R′i, R−i)(i) = c(R′i, R−i)(i) we
have c(R)(i)Ric(R′i, R−i)(i). By the inductive hypothesis c[ν] is non-bossy
and c[ν](R)(i) = c[ν](R′i, R−i)(i) implies c[ν](R) = c[ν](R
′i, R−i). Conse-
quently c(R)(i) = c(R′i, R−i)(i) (which holds if and only if c[ν](R)(i) =
c[ν](R′i, R−i)(i)) implies c(R) = ν ∪ c[ν](R) = ν ∪ c[ν](R
′i, R−i) = c(R′i, R−i).
Case II: The only ν ∈ N (cν)(R) is such that i ∈ Nν . Suppose there
existed a cycle not involving i at ∅ under c(R′i, R−i). Since the preferences of
all agents other than i are identical under R and (R′i, R−i) this cycle would
also exist at ∅ under c(R) contradicting the assumption that {ν} = N (c∅)(R).
Since there must be some cycle at ∅ under c(R′i, R−i), i is part of such a
cycle and his match c(R′i, R−i)(i) is the house that he points to at ∅ under
c(R′i, R−i). If c(R′i, R−i)(i) = c(R)(i) then the same cycle forms at ∅ under
c(R′i, R−i) and under c(R) and we obtain c(R′i, R−i) = ν∪c[ν](R) = c(R). So
c is non-bossy. Since c(R)(i) is the Ri-best house among all houses that i may
point to at ∅ under c, c(R)(i)Ric(R′i, R−i)(i) must hold and c is strategyproof.
To see that c is Pareto optimal fix any ν ∈ N (c∅)(R). By (C1) ν(i)RiH
is violated for at most one agent in Nν . This agent is a broker. Let c∅(hb) =
(ib, b). Since the brokered house hb is the only house that the broker may
not point Rib : hb ν(ib) must hold. Since ib ∈ Nν , ν(i∗) = hb holds for some
other agent i∗ ∈ Nν . Since i∗ is an owner hbRi∗H holds. So we cannot make
ib any better off without making i∗ worse off, implying that we cannot make
any agent in Nν any better off without making some agent in N worse off.
Since c(R) = ν ∪ c[ν](R) and since c[ν] is Pareto optimal by the inductive
hypothesis, there is no µ that Pareto dominates c(R).
6.3 Uniqueness
Let c′ be another control rights structure that defines the same mechanisms
as c. If c and c′ are both braids derived from different avoidance matchings
or if only one of the two mechanisms is a braid, then c and c′ define different
16
mechanisms. So suppose that according to c∅ and c′∅ there is at most one
brokered house. Fix any e ∈ H. If c∅(e) = (j, o) and c′∅(e) = (j′, o) holds for
some j 6= j′, then c(R∗)(e) = j 6= j′ = c′(R∗)(e) holds for R∗i : e for all i (in
each case the owner of the universally most preferred house e points to it in
the first round).
Now let c∅(e) = (j, b) 6= c′∅(e). If c′∅(e) = (j′, o) (possibly j = j′), then
(C2) implies that c∅(g) = (k, o) holds for some g 6= e and k /∈ {j, j′}. Letting
Rgi : e g for all i we obtain c(Rg)(k) = e and c′(Rg)(j′) = e. If c′∅(e) = (j′, b)
with j 6= j′, then c(Rg)(j) = g = c′(Rg)(j′) holds. In sum we obtain that
c∅ = c′∅ must hold for c and c′ to define the same mechanism. Since c∅ = c′∅,
ν is a direct c-successor of ∅ if and only if it is a direct c′-successor of ∅. The
hypothesis of the induction implies that c[ν] is identical to c′[ν] for any such
ν ∈ N (c∅).
6.4 A collection of arguments
Fix an arbitrary good mechanism M , a profile of preferences R, and a devi-
ation R′i. Let M(R)(i) = e. The following arguments are used throughout
the next sections. Strategy proofness implies that nothing changes for agent
i when he ranks e = M(R)(i) at least as high under R′i as under Ri:
SP-I If eRih⇒ eR′ih for all h ∈ H, then M(R′i, R−i)(i) = e.
Since M is non-bossy we additionally obtain:
SP-NB If eRih⇒ eR′ih for all h ∈ H, then M(R′i, R−i) = M(R).
If R′i and Ri differ only on the relative ranking of two houses, we obtain:
SP-II If eRig, gR′ie and R′i coincides with Ri on H \ {e, g} for some g 6= e,
then M(R′i, R−i)(i) ∈ {e, g}.
In combination with Pareto optimality the preceding observation yields
SP-PO If eRig, gR′ie and R′i coincides with Ri on H \ {e, g} for some g 6= e
and if M(R) is not Pareto optimal at (R′i, R−i), then M(R′i, R−i)(i) = g.
Finally I also use Lemma 5 from Papai [14]
L5-Papai If j is such thatM(R)(j)RiM(R)(i) thenM(R)(j)RjM(R′i, R−i)(j).
17
6.5 The Definition of c∅
Define a function c∅ : H → N × {o, b} as follows. Fix some house e and for
any g 6= e define Rg such that Rgi : e g holds for all i. If M(Rg)(i) = e holds
for all Rg with g 6= e then let c∅(e) = (i, o). If not let c∅(e) = (j, b) where j is
such that M(Rg)(j) = g. The goal of the current section is to show that c∅ is
well-defined. Lemma 2 shows that if M(Rg)(i) = e holds for a particular Rg
with Rg : e g then M(Rg)(i) = e holds for any Rg, implying that we either
have c∅(e) = (·, o) or c∅(e) = (·, b), not both. If c∅(e) = (·, o) then there is
a unique agent i for whom M(Rg)(i) = e holds for all g and c∅(e) = (i, o)
is well-defined. Finally Lemma 3 shows that there exists a unique agent ibwho obtains the second best house in any profile Rg when c∅(e) = (·, b) and
c∅(e) = (ib, b) is well-defined. Consequently c∅ is a well-defined function.
Lemma 2 Fix a set of m+ 2 different houses {h1, · · · , hm, e, g}. Let R and
R∗ be such that Ri : h1 · · ·hm e g and R∗i : h1 · · ·hm e g for all i. Then
M(R)(i∗) = e implies M(R∗)(i∗) = e.
Proof Suppose we had M(R)(2) = e and M(R′1, R−1)(3) = e for some
R′1 : h1 · · ·hm e g. Since M(R) 6= M(R′1, R−1) and since M is non-
bossy, g′ : = M(R)(1) must differ from h′ : = M(R′1, R−1)(1). Since R1 :
h1 · · ·hm e g and R′1 : h1 · · ·hm e g strategy proofness implies that g′, h′ /∈{h1 · · ·hm, e, g}. By SP-NB it is w.l.o.g to assume that R1 : h1 · · ·hm e g g′ h′and R′1 : h1 · · ·hm e g h′ g′. Letting M(R)(i) = g observe that i 6= 1, 2 since
M(R)(2) = e 6= g 6= g′ = M(R)(1). If i = 3 a violation of L5-Papai
would arise, in that case we would have g = M(R)(3)R1M(R)(1) = g′ and
e = M(R′1, R−1)(3)R3M(R)(3). Assume that i = 4 and R4 : h1 · · ·hm e g g′
(which is by SP-NB, w.l.o.g.). Define R′i : h1 · · ·hm g e for i = 2, 3 and
R′4 : h1 · · ·hm e g′ g to coincide with Ri on H \ {h1, · · · , hm, e, g} and re-
spectively H \ {h1, · · · , hm, e, g, g′}.SP+NB implies
(A) : M(R) = M(R′3, R−3).
SP-PO together with M(R′3, R−3)(2) = e imply
(B) : M(R′{2,3}R{1,4})(2) = g.
18
SinceM(R′{2,3}, R{1,4})(4) 6= g, SP-NB impliesM(R′{2,3}, R{1,4}) = M(R′−1, R1)
and particularly
(C) : M(R′−1, R1)(2) = g.
SP-PO and (A) yield M(R′{3,4}, R{1,2})(4) = g′ which together with SP-NB
implies
(D) : M(R′{3,4}, R{1,2}) = M(R′−2, R2).
SP+NB also impliesM(R′1, R−1) = M(R′{1,2}, R{3,4}), particularlyM(R′{1,2}, R{3,4})(3) =
e, which - given SP-II - implies
(E) : M(R′−4, R4)(3) ∈ {e, g}.
SP-PO and the initial assumption that M(R′1, R−1)(3) = e imply
(F ) : M(R′{1,3}, R{2,4})(3) = g.
SinceM(R′{1,3}, R{2,4})(4) 6= g, SP-NB impliesM(R′{1,3}, R{2,4}) = M(R′−2, R2).
So we may extend (D) to
(D′) : M(R′{3,4}, R{1,2}) = M(R′−2, R2) = M(R′{1,3}, R{2,4}).
It follows from (C) and SP-II that M(R′{3,4}, R{1,2})(2) ∈ {e, g}. Given (D′)
and (F )M(R′{3,4}, R{1,2})(2) cannot equal g, and we obtainM(R′{3,4}, R{1,2})(2) =
e. Using (D′) once again we obtain M(R′{1,3}, R{2,4})(2) = e. By SP-II
(G) : M(R′−4, R4)(2) ∈ {e, g}.
Together (E) and (G) imply that either
(H) : M(R′−4, R4)(2) = e and M(R′−4, R4)(3) = g or
(I) : M(R′−4, R4)(2) = g and M(R′−4, R4)(3) = e hold.
Suppose that (H) did hold. Given (B) we would then have M(R′−4, R4)(2) =
eR′1M(R′−4, R4)(1) /∈ {e, g} and M(R′{2,3}, R{1,4})(2) = gR′2M(R′−4, R4)(2) =
e a contradiction to Papai-L5. So (I) must hold. Applying SP-NB to (I) we
obtainM(R′−4, R4) = M(R′{1,2}, R{3,4}) and in particularM(R′{1,2}, R{3,4})(2) =
g. SP-II implies that M(R′1, R−1)(2) ∈ {e, g}. Our initial assumption that
19
M(R′1, R−1)(3) = e, then implies M(R′1, R−1)(2) = g. A contradiction
with Papai-L5 arises since M(R′1, R−1)(2) = gR′1M(R′1, R−1)(1) = h′ and
M(R)(2) = eR2M(R′1, R−1)(2) = g.
In sum we obtain that M(R′1R−1)(2) = e holds for any R′1 : h1 · · ·hm e g.
Since 1 was chosen arbitrarily, we can inductively obtain the above argument
to obtain that e = M(R)(2) = M(R∗1, R−1)(2) = M(R∗{1,2}, R−{1,2})(2) =
· · · = M(R∗)(2). �
Lemma 3 Let M(Rg)(1) = e = M(Rg′)(2) for some Rg, Rg′. Then there
exists an agent ib such that M(Rh)(ib) = h for any h 6= e.
Proof Fix Rg′
i : e g′ g for all i and let M(Rg′)(ib) = g′. To show that
M(Rg)(ib) = g let let Rg′
i : e g′ g coincide with Rgi on H \ {g′} for all i.
Since Rg′
i and Rg′
i agree on e g′ g, Lemma 2 implies M(Rg′)(ib) = g′ =
M(Rg′)(ib). Inductively switching Rg′
i to Rgi for all i 6= ib SP-NB implies
that M(Rg′
ib, Rg−ib) = M(Rg′). Let Rg
ib: e g g′ coincide with Rg
ibon H \ {g′}.
By SP-II M(Rgib, Rg−ib)(ib) ∈ {g, g
′}. If M(Rgib, Rg−ib)(ib) = g′, then non-
bossiness implies that M(Rgib, Rg−ib) = M(Rg′
ib, Rg−ib) = M(Rg′). Lemma 2
and M(Rg′)(2) imply M(Rg′)(2) = e and therefore M(Rgib, Rg−ib)(2) = e. On
the other hand Lemma 2 and M(Rg)(1) = e imply M(Rgib, Rg−ib)(1) = e, a
contradiction. So we must have M(Rgib, Rg−ib)(ib) = g. SP-I then implies that
M(Rg)(ib) = g.
Using the same arguments as above switching the roles of g and g′ (using
that M(Rg)(ib) = g holds for any Rg, particularly a profile in which all agents
rank g′ third) we obtain that M(Rg′)(ib) = g′ holds for any Rg′ . Now fix any
h 6= e. Apply the above arguments to Rh and Rh′ with Rh′i : e h′ h for
all i where h′ = g′ if M(Rh)(e) = 2 and h′ = g otherwise to obtain that
M(Rh)(ib) = h holds for all Rh. �
6.6 Properties of c∅
In the following Lemmas 4, 5, and 6 show that c∅ satisfies a range of proper-
ties. Lemma 4, which is used in the proofs of nearly all upcoming Lemmas,
shows that {(i, e), (j, g)} ⊂ M(R′) holds for any R′ with R′i : e and R′j : g
20
or R′j : e g if c∅(e) = (j, b) and e = M(Rg)(i). Lemma 5 shows that i is
matched to e under M(R) if he owns e according to c∅ (c∅(e) = (i, o)) and if
i prefers e to all other houses (Ri : e). Lemma 6 shows that any broker only
brokers a single house and does not own any house: c∅(e) = (i, b) implies
c∅(h) 6= (i, ·) for any h 6= e. If all agents rank some e with c∅(e) = (i, b) at
the top, then the owner of the second most preferred house of i is matched to
e. The lemmas in the current section condense the lemmas of the preceding
section. The lemmas of the preceding section are not (directly) used after
the current section.
Lemma 4 Let c∅(e) = (1, b). If M(Rg)(2) = e holds for some Rg, then
M(R)(2) = e and M(R)(1) = g holds for all R such that R2 : e and either
R1 : g or R1 : e g.
Proof Fix any profile R such that R2 : e and either R1 : g or R1 : e g. For
all i let Rgi : e g and Ri coincide on H \ {e, g}. Lemma 2 and the assumption
that M(Rg)(2) = e implies M(Rg)(2) = e. Lemma 3 and the and c∅(e) =
(1, b) imply M(Rg)(1) = g . Inductively apply SP-NB, dropping the houses
e and g in the rankings of all agents i 6= 1, 2 to obtain M(Rg1,2, R−{1,2}) =
M(Rg). Drop house g in agent 2’s ranking and house e in agent 1’s ranking
if R1 : g to obtain M(Rg1,2, R−{1,2}) = M(R) in particular M(R)(1) = g and
M(R)(2) = e. �
Lemma 5 If c∅(e) = (1, o) and R1 : e, then M(R)(1) = e.
Proof Showing M(R)(1) = e if Ri : e holds for all i is sufficient. To see
this fix any R◦ with R◦1 : e let Ri : e and R◦i coincide on H \ {e} for all i. If
M(R)(1) = e, then M(R◦)(1) = e follows from the inductive application of
SP-NB.
Now suppose M(R)(1) 6= e held for some Ri : e for all i and R1 : e g. For
all i, let Rgi coincide with Ri on H \ {g}, M(Rg)(j) = g and M(R)(j) = g′.
If gRjg′ then SP-NB yields M(R) = M(Rg
j , R−j). Given that Lemma 4 im-
plies M(Rgj , R−j) = M(Rg) we obtain M(R) = M(Rg) and the contradiction
e 6= M(R)(1) = M(Rg)(1) = e since c∅(e) = (1, o). So g′Rjg and g′ 6= g
must hold. SP-NB implies M(R) = M(R′j, R−j) for R′j : e g′ g. To ob-
tain a contradiction consider agent 1’s match under M(Rg′
1 , R′j, R−{1,j})(1).
21
Let Rg′
i : e g′ coincide with R′j and with Ri for i 6= 1, j on H \ {g′}and let R∗1 : g′. Since c∅(e) = (1, o) we have M(Rg′)(1) = e. SP-PO
implies M(R∗1, Rg′
−1)(1) = g′, the inductive application of SP-NB implies
M(R∗1, R′j, R−{1,j})(1) = g′. Combining the last statement with SP-PO yields
M(Rg′
1 , R′j, R−{1,j})(1) ∈ {e, g′}. If M(Rg′
1 , R′j, R−{1,j})(1) = e then agent 1
gains by misrepresenting his preference at (R′j, R−j) and M is not strate-
gyproof. If M(Rg′
1 , R′j, R−{1,j})(1) = g′ then M(Rg′
1 , R′j, R−{1,j})(j) 6= g′ and
SP-NB imply M(Rg′
1 , Rgj , R−{1,j}) = M(Rg′
1 , R′j, R−{1,j}) which yields a con-
tradiction to Lemma 4 which requires that M(Rg′
1 , Rgj , R−{1,j})(1) = e 6= g′.
�
Lemma 6 Let c∅(e) = (1, b). Then
a) c∅(h) 6= (1, o) holds for all h ∈ H.
b) c∅(h) 6= (1, b) holds for all h ∈ H \ {e}.c) If c∅(g
∗) = (j, o) then M(Rg∗)(j) = e.
Proof Since c∅(e) = (1, b), it is w.l.o.g to assume that M(Rg)(2) = e,
M(Rg′)(3) = e, for some g 6= e 6= g′.
a) Since M(Rg)(e) 6= 1 even though Rg1 : e, Lemma 5 implies c∅(e) 6=
(1, o). Suppose c∅(g) = (1, o) for some g 6= e. Let R∗2 : g e. SP-PO implies
M(R∗2, Rg−2)(2) = g. Since c∅(g) = (1, o) Lemma 5 and strategyproofness im-
ply M(R∗2, Rg−2)(1)Rg
1g. According to Rg1 only e and g are (weakly) preferred
to g. Since g is matched to 2 under (R∗2, Rg−2) we obtain M(R∗2, R
g−2)(1) =
e. SP-I then implies M(Rg′
1 , R∗2, R
g−{1,2})(1) = M(R∗2, R
g)(1) = e. But
Lemma 4 together with M(Rg′)(e) = 3, Rg′
1 : e g′, and Rg3 : e implies
M(Rg′
1 , R∗2, R
g−{1,2})(3) = e.
b) Suppose c∅(g) = (1, b) held for some g 6= e. So there exists some
house h such that M(R◦)(j) = g, with j /∈ {1, 2} and M(R◦)(1) = h where
R◦i : g h holds for all i. For all i let R∗i : e g h coincide with R◦i on H \ {e}.The assumptions M(Rg)(2) = e and M(Rg)(1) = g together with Lemmas
2 and 3 imply that M(R∗)(2) = e and M(R∗)(1) = g. Given that neither 1
nor j 6= 2 is matched to e under M(R∗) the inductive application of SP-NB
implies M(R∗)(1) = M(R◦{1,j}, R∗−{1,j})(1) = g. But Lemma 4 requires the
contradiction M(R◦{1,j}, R∗−{1,j})(j) = g.
22
c) The definition of c∅ and the assumption c∅(e) = (1, b) implyM(Rg)(1) =
g. Since c∅(g∗) = (j, o) Lemma 5 implies that agent j is matched to a house
M(Rg∗)(j)Rg∗
j g∗. Since e is the only house other than g∗ that satisfies this
relation M(Rg∗)(j) = e must hold. �
6.7 Braids, (C1), (C2), and (C3)
In Lemma 7 I show that M is a braid if c∅ calls for at least two houses to
be brokered, implying that c∅ satisfies (C1). If there is at most one brokered
house according to c∅, let i, e be such that c∅(e) = (i, b). Part a) of Lemma
6 then implies c∅(h) 6= (i, o) as required by (C3). Finally part c) of Lemma
6 together with the definition of c∅ implies that there must be at least two
owners under c∅ for there to be a broker under c∅ as required by (C2). In
sum, c∅ satisfies (C1), (C2) and (C3).
Lemma 7 Let c∅(e) = (1, b) and c∅(g) = (k, b) for e 6= g. Then | H |= 3
and M is a braid.
Proof W.l.o.g. assume that M(Rg)(2) = e = M(Rg′)(3) for some Rg, Rg′ .
Claim 1: Fix R∗ with R∗i : g e for all i. Then M(R∗)(2) = g must hold.
Fix Rgi : e g to coincide with R∗i on H \ {g} for all i. Inductively dropping
house e in the rankings Rgi of all agents i 6= 2, SP-NB yields M(Rg) =
M(Rg2, R
∗−2). If M(Rg
2, R∗−2) = M(R∗) then 2 = k and M(R∗)(1) = g must
hold. Lemma 4 implies M(R∗{1,2}, Rg−{1,2})(1) = g. SP-II applied to agent 1’s
choice yields M(R∗2, Rg−2)(1) ∈ {e, g}. Since M(R∗2, R
g−2)(2) = g holds due
to SP-PO and M(Rg)(2) = e M(R∗2, Rg−2)(1) must equal e. SP-NB implies
M(R∗2, Rg′
1 , Rg−{1,2}) = M(R∗2, R
g−2) in particular M(R∗2, R
g′
1 , Rg−{1,2})(1) = e.
Lemma 4 and the assumption M(Rg′)(3) = e imply M(R∗2, Rg′
1 , Rg−{1,2})(3) =
e, a contradiction. So M(Rg2, R
∗−2) must differ from M(R∗). Non-bossiness
and SP-II then imply e = M(Rg2, R
∗−2)(2) 6= M(R∗)(2) = g.
Claim 2: c∅(g) = (k, b), so k = 3.
Since c∅(e) = (1, b) part b) of Lemma 6 implies that k 6= 1. Since M(R∗)(2) =
g as established in Claim 1 k 6= 2. Claim 2 follows since the assumption that
23
k > 3 together with Lemma 4, M(Rg′)(1) = e and M(R∗)(2) = e leads to
the contradiction that
M(R∗{2,k}, Rg′
−{2,k})(k) = M(R∗{2,k}, Rg′
−{2,k})(1) = e.
Claim 3: There is no h ∈ H and j > 3 such that c∅(h) = (j, o).
Suppose c∅(h) = (j, o) held for some j > 3 and h ∈ H. Fix some Rh. Since
c∅(e) = (1, b), M(Rh)(1) = h holds by the definition c∅. Since c∅(h) = (j, o)
part c) of Lemma 6 implies M(Rh)(j) = e. By Lemma 4 together with
M(Rh)(j) = e, M(R∗)(2) = g, M(R∗)(3) = e (as established in Claims 1
and 2) we obtain the contradiction
M(Rh{1,j}, R
∗−{1,j})(j) = M(Rh
{1,j}, R∗−{1,j})(3) = e.
Claim 4: There is no h ∈ H and j > 3 such that c∅(h) = (j, b).
Suppose c∅(h) = (j, b) held for some j > 3 and h ∈ H. Fix Rαi : h e
for all i and let M(Rα)(j′) = h. Mutatis mutandis Claim 1 implies that
M(Rh)(j′) = e. Since c∅(e) = (1, b), j′ cannot equal 1. If j′ 6= 2 then Lemma
4 and M(Rg)(2) = e imply
M(Rg{1,2}, R
α−{1,2})(2) = M(Rg
{1,2}, Rα−{1,2})(j) = e
a contradiction since j > 3. If j′ = 2 a similar contradiction is obtained
replacing agent 2 with 3 and house g with g′.
Claim 5: c∅(g′) = (2, b) and H = {e, g, g′}.
If c∅(h) = (i, ·) holds for some h ∈ H then i ≤ 3 follows from Claims 3 and 4.
Parts a) and b) of Lemma 6 applied to c∅(e) = (1, b) and c∅(g) = (3, b) (which
follow from the initial assumption and Claim 2) yield c∅(g′) 6= (1, ·) and
c∅(g′) 6= (3, ·) Since M(Rg′)(2) /∈ {e, g′} c∅(g′) 6= (2, o). The only remaining
option is c∅(g′) = (2, b). Since for any i ∈ {1, 2, 3} there is a house h ∈
{e, g, g′} such that c∅(h) = (i, b) parts a) and b) of Lemma 6 imply that
c∅(h) 6= (i, ·) holds for all h /∈ {e, g, g′} and i ≤ 3. In sum we obtain
H = {e, g, g′}.
Claim 6: M(R) = B(R).
24
I first show that M and B match the the top ranked houses to the same
agents if R is such that Ri = Rj for all i, j. For R ∈ {Rg, Rg′ , R∗} the initial
assumption as well as Claims 2 and 3 determine the agents that are matched
with the two top ranked houses. Since c∅(g) = (3, b) there must exist a house
h /∈ {e, g} and an agent j /∈ {2, 3} such that M(R◦)(j) = g for R◦i : g h for
all i. Since there are only three houses h must equal g′. Moreover j = 1,
since otherwise Lemma 4 implies the contradiction
M(Rg{1,2}, R
◦−{1,2})(1) = M(Rg
{1,2}, R◦−{1,2})(3) = g.
Let Ri : g′ g and Ri : g′ e for all i. By Claim 4 c∅(g′) = (2, b) and we have
M(R)(2) = g and M(R)(2) = e. Mutatis mutandis Claim 1 then implies
that M(R)(1) = g′ = M(R)(3).
At Rg ω′ and ω′′ are both Pareto optimal. Since at least two agents rank
e at the top under Rg, since ω(e) = 1 and g = ω′(1)Rg1ω′′(1) = g′′ B(Rg) = ω′
must hold, in particular we have B(Rg)(1) = g and B(Rg)(2) = e, so B and
M match the two top ranked houses under Rg to the same agents. The proof
this statement holds for all R ∈ {Rg′ , R∗, R◦, R, R} follows from the same
arguments mutatis mutandis.
To see thatM(Rg)(3) = g′, observe that SP-NB impliesM(Rg) = M(R∗1, Rg2, R
◦3).
Lemma 4 together with M(R◦)(1) = g and M(R◦)(3) = g′ (as established
in the preceding paragraph) then implies M(R∗1, Rg2, R
◦3)(3) = g′. Mutatis
mutandis the same arguments prove that M(R)(i) ∈ {e, g, g′} holds for
all R ∈ {Rg′ , R∗, R◦, R, R} and i ≤ 3. So M(R) = B(R) holds for all
R ∈ {Rg, Rh, R∗, R◦, R, R}.Fix R with Rj : ω′(j), ω′(j′)Rj′ω
′′(j′) and ω(j′) = ω′(j) (or Rj : ω′′(j),
ω′′(j′)Rj′ω′(j′) and ω(j′) = ω′′(j)) for some j 6= j′. Let j′ = 1 and j = 2,
so R2 : e, ω′(1) = gR1g′ = ω′′(1) and ω(1) = ω′(2) = e. We know from (**)
in the proof of Lemma 1 that B(R) = ω′. The arguments in the preceding
two paragraphs yield M(Rg) = B(Rg) = ω′. If Rg 6= R, then the inductive
application of SP-NB implies M(R) = M(Rg) = B(Rg) = B(R). For all
other combinations of agents j, j′ the M(R) = B(R) is established mutatis
mutandis.
Fix R with Rj : h with h 6= ω′(j), ω′(j′)Rj′ω′′(j′), ω(j′) = ω′(j) and
Rj′′ : ω′(j) where {j, j′, j′′} = {1, 2, 3}. Let j′ = 1 and j = 2, so R2 : h,
25
ω′(1) = gR1g′ = ω′′(1) and ω(1) = ω′(2) = e and R3 : e. Let R′2 :
e. Since M(Rg′
1 , R−1)(1) = g′, SP-I implies that M(R)(1) 6= e. Since
M(Rg2, R−2)(2) = g and since M is strategy proof M(R)(2) ∈ {e, g, g′}. Since
M is Pareto optimal M(R)(2) 6= e. So M(R)(3) = e. Since M(R′2, R−2)(2) =
e agent 2 must prefer M(R)(2) to e. Since we only specified agent 2 ranks
h 6= e at the top under R2 this would have to apply in particular if agent 2
ranks e in second place, implying M(R)(2) = h. Since M(Rg′
1 , R−1)(1) = g′
strategyproofness implies that agent 1 is matched with some house at R.
Given that M(R)(2) = h and M(R)(3) = e, M(R)(1) is the one remaining
house in {e, g, g′}. The case that Rj : h with h 6= ω′′(j), ω′′(j′)Rj′ω′(j′),
ω(j′) = ω′′(j) and Rj′′ : ω′′(j) where {j, j′, j′′} = {1, 2, 3} and all alternative
combinations of agents j, j′ and a house h can be dealt with with the same
arguments mutatis mutandis.
Only once case remains to be considered R is such that Ri : ω(i) for all i
(as in this case i is the only agent to rank ω(i) at the top). The two preceding
paragraphs imply that for any i ≤ 3 there exists a R′i : h with ω(i) 6= h such
that M(R′i, R−i)(i) ∈ {e, g, g′}. Since M is strategyproof M(R)(i) 6= ∅ must
hold for all i ≤ 3. But there is only one Pareto optimum at R with this
feature M(R) = ω and we also obtain M(R) = B(R) = in this last case. �
6.8 Pointing
Assume that M is not a braid and fix some R◦. Here I show that the outcome
M(R◦) is consistent with submatching achieved under R◦ in the first round of
any trading and braiding mechanism with c∅ the control rights function at ∅.Recall the definitions ofN (c∅) andN (c∅)(R) in Section 6.2 as the set of direct
c-successors to ∅ and the subset of direct c-successors to ∅ that are reachable
under the profile of preferences R. Since only c∅ is used in determining
the set of direct c-successors to ∅, N (c∅) and N (c∅)(R) describe the same
sets for any trading and braiding mechanism c that prescribes c∅ at ∅. In
Lemma 8 I show that ν ⊂ M(R◦) holds for any ν ∈ N (c∅)(R). In Lemma 9
I show that the calculation of M(R◦) can be split into the calculation of a
submatching ν ∈ N (c∅)(R) and the outcome of a well-defined submechanism.
The submechanism inherits the property of being good from M and can by
26
the inductive hypothesis be represented as a trading and braiding mechanism
c[ν]. For the next two Lemmas fix a submatching ν ∈ N (c∅)(R◦). W.l.o.g.
assume Nν = {1, · · · ,m} and c∗∅(hi) = (i, ·) for all i ≤ m.
Lemma 8 Any submatching that arises out of matching one cycle under c∅at R◦ is part of the outcome M(R◦), ν ∈ N (c∅)(R
◦) implies ν ⊂M(R◦).
Proof Case 1: m = 1. So ν = {(i, h)} holds for some i, h and c∅(h) = (i, ·)must hold since h points to i at ∅. Since i may only point to h if c∅(h) 6= (i, b),
c∅(h) = (i, o) must hold. Since i, as an owner may point to any house, R◦i : h
must hold. Lemma 5 then implies M(R◦)(i) = h and therefore ν ⊂M(R◦).
Case 2: m > 1 and c∅(hi) = (i, o) for all i ≤ m. W.l.o.g assume that
ν(i) = hi+1 for all i < m and ν(m) = h1. For all i ≤ m let R∗i : ν(i) hicoincide with Ri on H \ {hi} and let R∗i = R◦i for i > m. Suppose we had
M(R∗)(i∗) 6= ν(i∗) for at least one i∗ ≤ m, say i∗ = m. Since c∅(hm) = (m, o)
Lemma 5 implies M(R′m, R∗−m)(m) = hm for R′m : hm. SP-II implies that
M(R∗)(m) ∈ {h1, hm}. The assumption M(R∗)(m) 6= ν(m) = h1 then
implies M(R∗)(m) = hm. So M(R∗)(m − 1) differs from ν(m − 1) = hm.
Inductively applying these arguments to all agents in the cycle we obtain
that M(R∗)(i) = hi for all i ≤ m. This contradicts the Pareto optimality
of M since any i ≤ m strictly prefers ν(i) to hi. So ν ⊂ M(R∗) must hold.
Inductively applying SP-NB to drop hi in the rankings of all agents i ≤ m
we obtain M(R∗) = M(R◦).
Case 3: c∅(hm) = (m, b). Since m, as a broker, may not point to hmwe have m > 1. As in the preceding case assume that ν(i) = hi+1 for all
i < m and ν(m) = h1. Define R∗m−2 : hm−1 h1, R∗m : hm h1 hm−1 and
R∗i : ν(i) hi for all other i ≤ m. For all preference statements that have not
been explicitly mentioned let R∗ and R◦ coincide. Under M(R′m−2, R∗−(m−2))
with R′m−2 : h1 the owners {1, · · · ,m − 2} form a pointing cycle. Applying
the result of Case 2 we obtain M(R′m−2, R∗−(m−2))(m−2) = h1. Lemma 4 and
part c) of Lemma 6 imply that M(R′m, R∗−m)(m) = hm−1 for R′m : hm hm−1.
Strategyproofness implies M(R∗)(m − 2)R∗m−2h1, M(R∗)(m)R∗mhm−1, and
M(R∗)(m) 6= hm. So then agents {m−2,m} must be matched to the houses
{h1, hm−1} under R∗. Pareto optimality requires that M(R∗)(m−2) = hm−1and M(R∗)(m) = h1. The last observation in turn implies that M(R∗)(1) 6=
27
h1. Since c∅(h1) = (1, o), Lemma 5 together with strategyproofness implies
that M(R∗)(1)R∗1h1. So M(R∗)(1) must equal h2. We can now inductively
apply the same arguments to all other agents to obtain that ν ⊂ M(R∗).
Inductively applying SP-NB to drop h1 in R∗m−2, hm−1 and hm in R∗m, and
hi in all other rankings R∗i with i ≤ m we obtain M(R∗) = M(R◦). �
Keeping ν ∈ N (c∅)(R) fixed as above, let R be the restriction of R to
the agents N ν over houses Hν . Let R be the set of all such restrictions.
Define a trading and braiding mechanism c[ν] that maps any R ∈ R to
c[ν](R) : = M(R) \ ν.
Lemma 9 The mechanism c[ν] is well-defined.
Proof Fix another profile R∗ such that ν ∈ N (c∅)(R∗). Only the pref-
erences of agents i ≤ m matter for the formation of ν under M(R∗) and
M(R◦). So the mechanisms M∗ and M◦ that respectively map any R ∈ Rto M∗(R) : = M(R∗Nν , R−Nν ) \ ν and M◦(R) : = M(R◦Nν , R−Nν ) \ ν are
well-defined.
To see that M∗ = M◦, fix some some R ∈ R and note
ν ∪M∗(R) = M(R∗Nν , R−Nν ) =
M(R◦1, R∗{2,··· ,m}, R−Nν ) = · · · = M(R◦{1,··· ,m}, R
∗m, R−Nν ) =
M(R◦Nν , R−Nν ) = ν ∪M◦(R).
The first and the last equality follow from the definitions ofM∗ andM◦. Since
ν ∈ N (c∅)(R∗) and ν ∈ N (c∅)(R
◦), R∗i : ν(i), R◦i : ν(i) must hold for all own-
ers i ≤ m (that is all agents i ≤ m with c∅(hi) = (i, o)) and νi(i)R∗iH \ {hi},
νi(i)R◦iH \ {hi} must hold for a broker i ≤ m (if there is one) where c∅(hi) =
(i, b) and hi is the house he brokers. So the cycle also arises under c∅ at any
of the intermediate profiles (R◦1, R∗{2,··· ,m}, R−Nν ) · · · (R◦{1,··· ,m−1}, R∗m, R−Nν ).
Lemma 8 implies that ν ⊂M(R◦1, R∗{2,··· ,m}, R−Nν ), · · · , ν ⊂M(R◦{1,··· ,m}, R
∗m, R−Nν ).
Given that agent i ≤ m is matched to ν(i) for any (R◦1, R∗{2,··· ,m}, R−Nν ) · · · (R◦{1,··· ,m−1}, R∗m, R−Nν )
the non-bossiness of M the implies the intermediate equalities.
To see that M◦ is strategy proof fix an agent i ∈ Nν and a deviation R′i
for agent i. Let R′i and R be such that R′i and R are the restrictions of R′i
28
and R to Hν , N ν . The definition of M◦ together with the strategyproofness
of M then imply
M◦(R)(i) = M(R◦Nν , R−Nν )(i)RiM(R◦Nν , R′i, R−(Nν\{i}))(i) = M◦(R
′i, R−i)(i).
To see that M◦ is non-bossy assume that M◦(R)(i) = M◦(R′i, R−i)(i). Since
i /∈ Nν we obtain M(R◦Nν , R−Nν )(i) = M(R◦Nν , R′i, R−(Nν\{i}))(i). The non-
bossiness of M then implies M(R◦Nν , R−Nν ) = M(R◦Nν , R′i, R−(Nν\{i})) and
therefore
M◦(R) = M(R◦Nν , R−Nν ) \ ν = M(R◦Nν , R′i, R−(Nν\{i})) \ ν = M◦(R
′i, R−i).
Finally M◦ is Pareto optimal since M(R◦Nν , R−Nν ) = ν∪M◦(R) is Pareto
optimal for any R. In sum M◦ is a good mechanism. By the hypothesis of
the induction M◦ can be represented as a trading and braiding mechanism
c[ν]. �
6.9 Requirements that link control rights functions:
(C4), (C5) and (C6)
The definition of c∅ in Section 6.5 together with Lemma 9 on the submecha-
nisms c[ν] define a control rights structure c via cν◦ = c[ν]ν′ for any ν◦ = ν∪ν ′with ν ∈ N (c∅) and ν ′ c[ν]-relevant. Since a submatching ν◦ is c-relevant if
and only if it can be split into a submatching ν ∈ N (c∅) and a submatching
ν ′ that is c[ν]-relevant, c is defined on all c-relevant submatchings.
We know from Section 6.7 that c∅ satisfies (C1), (C2) and (C3). Since
c[ν] is a trading and braiding mechanism for any ν ∈ N (c∅), (C1), (C2) and
(C3) are satisfied for any c-relevant ν◦ 6= ∅. Moreover (C4), (C5) and (C6)
are satisfied by any pair of a c-relevant ν◦ 6= ∅ with a direct c-successor ν
of ν◦. In the following Lemma 10 I show that (C4), (C5) and (C6) are also
satisfied for ν◦ = ∅ and any ν ∈ N (c∅). Keep R◦, ν ∈ N (c∅)(R◦) and the
definition of R fixed as above.
Lemma 10 (C4), (C5) and (C6) hold for ∅ and ν ∈ N (c∅).
29
Proof (*) For any trading and braiding mechanism c◦ we have c◦∅(e) =
(i, o) holds if and only if c◦(R)(i) = e for all R such that Ri : e. If c◦∅(e) 6= (·, o)then c◦∅(e) = (i, b) if and only if c◦(Rg)(i) = g.
(C4) Let i∗ /∈ Nν be such that c∅(e) = (i∗, o). Fix any R with Ri∗ : e. Define
R such that Ri = R◦i for all i ∈ Nν and Ri∗ : e. Since i∗ owns* e Lemma
5 implies M(R)(i∗) = e. The results in the preceding section imply that
M(R) = ν ∪ c[ν](R), in particular M(R)(i∗) = e = c[ν](R)(i∗). By (*) i∗
owns e under c[ν]∅ = cν .
(C5) Let ib, j, j′ /∈ Nν be such that c∅(e) = (ib, b) and agents j and j′ respec-
tively own houses g and g′ under c∅. By the preceding paragraph on (C4), j
and j′ own g and g′ under cν . By Lemma 8 we have M(R◦NνRg−Nν )(ib) = g,
M(R◦NνRg−Nν )(j) = e = M(R◦NνR
g′
−Nν )(j′). By the preceding section we have
M(R◦NνRg−Nν ) = ν ∪ c[ν](R
g) and M(R◦NνR
g′
−Nν ) = ν ∪ c[ν](Rg′
). In sum we
get that c[ν](Rg)(j) = e = c[ν](R
g′
)(j′) and by (*) e is not owned at c[ν]∅.
Since M(R◦NνRg−Nν )(ib) = g = c[ν](R
g)(ib) (*) also implies that ib brokers e
under c[ν]∅ = cν .
(C6) Let c∅(g) = cν(g) = (j, o), c∅(e) = (ib, b) 6= cν(e) and ib /∈ Nν . Let
Ri = Rgi for all i ∈ N ν and let Ri = R◦i for all i ∈ Nν . Let R′i : h∗ e g for
some fixed house h∗ ∈ Hν . By Lemma 8 M(R)(ib) = g and M(R)(j) = e. If
ib did own e under c[ν]∅ we obtain M(R′ib , R−ib)(ib) = c[ν](R′ib, R−ib)(ib) = e
contradiction to strategyproofness.
Now suppose an agent k /∈ {ib, j} controls e at ∅ under c[ν]. If c[ν]∅(e) =
(k, b) we obtain a contradiction to strategyproofnes given that R′ib ranks g =
M(R)(ib) aboveM(R′ib , R−ib)(ib) = c[ν](R′ib, R−ib)(ib) since under c[ν](R
′ib, R−ib)
k, the (new) broker of e is matched with g and j the owner of g is matched
with e. If c[ν]∅(e) = (k, o) we obtain a similar contradiction to strategyproof-
ness: e = M(R)(j)R′jM(R′j, R−j)(j) = c[ν](R′j, R−j)(j) since c[ν](R
′j, R−j)(k) =
e. So cν(e) = (j, ·) must hold. Since j owns g at c[ν]∅ = cν , and since by
(C3) no broker owns a house, j must own e under c[ν]∅ = cν .
When c[ν]∅(e) = (j, o) the submatching {(j, e)} is c[ν]-relevant. To see
that ib must own g at this submatching fix any profile of preferences R for
the agents N ν \ {j} over the houses Hν \ {e} with Rib : g. Define R such
that Ri = R◦i for all i ∈ Nν , Rj : e, Rib : g and Ri is restriction of Ri to
30
Hν \ {e} for all i ∈ Nν \ {j}. Lemma 8 implies that M(R)(ib) = g. The
preceding section implies that M(R) = ν ∪ c[ν](R). Since c[ν] is a trading
and braiding mechanism with c[ν]∅(e) = (j, o) and since Rj : e we obtain
c[ν](R) = {(j, e)} ∪ c[ν ∪ {(j, e)}](R). In sum we have that M(R)(ib) = g =
c[ν∪{(j, e)}](R)(ib). So (*) implies that ib owns g under c[ν]{(j,e)} = cν∪{(j,e)}.
�
7 Conclusion
Let me highlight three features of the proof of Theorem 1. I started by
showing that the order of elimination of trading cycles does not matter in
the definition of trading and braiding cycles mechanisms. The freedom to
eliminate trading cycles in any order allows me to conveniently structure
the inductive proof that trading and braiding mechanisms are good. The
calculation of any outcome c(R) can in particular be split into finding one
direct c-successor ν of ∅ that is reachable under c(R) and the outcome of the
submechanism following ν: c[ν](R). Since any such c[ν] involves less than
n+ 1 agents the inductive hypothesis applies: so any such c[ν] is good. Only
very few cases need to be considered to show that c itself is good. These
cases all pertain to cycles that form at the start of a trading and braiding
mechanism c.
Secondly, while induction over the number of agents turns out to be a
major simplifier in the proof that any trading and braiding mechanism is
good, it plays only less important role in the proof that any good mecha-
nism can be represented as a trading and braiding mechanism. The core of
this part of the proof lies in definition of a control rights function c∅ and
in showing that braids can only arise with exactly three houses. Once this
groundwork has been laid the proof unfurls with far greater ease. The in-
ductive structure over the number of agents in the mechanism only matters
in the proof of Lemma 9. In this lemma I show that the calculation of any
M(R) (when M is not a braid) can be split into a first round of trading cycles
and a submechanism. This submechanism inherits the properties of Pareto
optimality, strategy proofness and non-bossiness from its parent M . Since
31
the submechanism involves fewer agents than its parent, the submechanism
has - by the inductive hypothesis - a representation as a trading and braiding
cycle.
Thirdly, all three assumptions: strategy proofness, Pareto optimality and
non-bossinss are repeatedly used in the proof that any good mechanism can
be represented as a trading and braiding mechanism. Already the very first
lemma in this proof (Lemma 2) requires all three properties. Given that strat-
egyproofness and Pareto optimality are better founded than non-bossiness as
principles of mechanism design one might wonder about a characterization
of the set of all strategy proof and Pareto optimal mechanisms. But, given
that my proof extensively relies on all three properties a simple extension
of the same proof to the grand set of all strategy proof and Pareto optimal
mechanisms is out of the question.
However, the representation of good mechanisms as trading and braiding
mechanisms can be used to construct a class of Pareto optimal and strategy
proof mechanisms. Simply modify control rights structures insofar as that
the inheritance of houses not only depends on submatchings ν but also on the
preferences of the matched agents. Such a modified control rights structure
c maps any combination of a c-relevant ν and a profile of preferences R
to a control rights function. Keeping (C1)-(C6) intact we would have to
additionally impose that a c-relevant ν and two profiles of preferences R and
R′ can only be mapped to two different control rights functions if Ri 6= R′iholds for some i ∈ Nν . A serial dictatorship in which the second dictator
depends on the first dictators preferences over houses he did not choose is
the simplest example of such a bossy mechanism. The present proof that any
trading and braiding mechanism is good can easily be extended to show that
any mechanism, defined through such a modified control rights structure, is
strategyproof and Pareto optimal. The question whether any Pareto optimal
and strategyproof mechanisms can be represented by such a modified control
rights structure awaits some new ideas and techniques of proof.
32
References
[1] Abdulkadiroglu, A. and Che Y.-K.: The Role of Priorities in Assigning
Indivisible Objects: A Characterization of Top Trading Cycles, (2012),
mimeo, Columbia University.
[2] Abdulkadiroglu , A. and T. Sonmez: “Random Serial Dictatorship and
the Core from Random Endowments in House Allocation Problems”,
Econometrica, 66, (1998), pp. 689 - 701.
[3] Abdulkadiroglu, A. and T. Sonmez: “House Allocation with Existing
Tenants”, Journal of Economic Theory, 88, (1999), pp. 233 - 260.
[4] Abdulkadiroglu, A. and T. Sonmez: “School Choice a Mechanism Design
Approach”, American Economic Review, 93, (2003), pp. 729 - 747.
[5] Bade, S.: “Serial dictatorship: the unique optimal allocation rule when
information is endogenous”, forthcoming Theoretical Economics
[6] Bade, S.: “Random Serial Dictatorship: the One and Only”, (2014),
mimeo, Royal Holloway College.
[7] Carroll, G.: “A General Equivalence Theorem for Allocation of Indi-
visible Objects,” Journal of Mathematical Economics, 51, (2014), pp.
163-177.
[8] Che Y.K., J. Kim and F. Kojima: “Efficient Assignment with Interde-
pendent Values”, (2014), mimeo, Stanford University.
[9] Ehlers, L. and A.Westkamp: “Strategy-proof Tie-breaking”, (2014),
mimeo, Universite de Montreal.
[10] Ekici, O.: “A General Fair and Efficient Discrete Resource Allocation:
A Market Approach,” (2012), mimeo, Ozyegin University.
[11] Knuth, D.: “An exact analysis of stable allocation”, Journal of Algo-
rithms, 20, (1996), pp. 431-442.
33
[12] Lee, T. and J. Sethuraman: “Equivalence results in the allocation of
indivisible objects: A unified view”, (2011), mimeo Columbia University.
[13] Ma, J.: “Strategyproofness and the Strict Core in a Market with In-
divisibilites”, International Journal of Game Theory, 23, (1994), pp.
75-83.
[14] Papai, S.: “Strategyproof Assignment by Hierarchical Exchange”,
Econometrica, 68, (2000), pp. 1403 - 1433.
[15] Pathak, P and J. Sethuraman.: “Lotteries in student assignment: An
equivalence result”, Theoretical Economics, 6, (2011), pp. 1-17.
[16] Pycia, M. and U. Unver: “Incentive Compatible Allocation and Ex-
change of Discrete Resources”, (2014), mimeo UCLA.
[17] Sonmez, T. and U. Unver: “House Allocation with Existing Tenants:
A Characterization”, Games and Economic Behavior, 69, (2010), pp.
425-445.
[18] Shapley, L. and H. Scarf: “On Cores and Indivisibility”, Journal of
Mathematical Economics, 1, (1974), pp. 23 - 37.
[19] Svensson L. G.: “Strategy-proof allocation of indivisible goods”, Social
Choice and Welfare 16, (1999), pp. 557-567.
[20] Velez, R.: “Consistent strategy-proof assignment by hierarchical ex-
change”, Economic Theory, 56, (2014), pp. 125 - 156.
34